Three Dimensional Reconstruction Of Structured Scenes Based On Vanishing Points

The goal of the computer vision is to recover visible surfaces of three dimensional objects from two dimensional images. After Marr vision theory framework established, 3D reconstruction technique has developed quickly and has widely used in robot navigation and obstacle recognition, virtual reality, remote sense analysis, vision measuring, architectural retrieval and other fields. Thanks to constrain of some key issues, such as image matching, many 3D reconstruction algorithms and systems are lack of robustness, and they are always designed for some special scene without generality. In daily life, objects and scenes that we come into contact with are always structured ones, so the measuring and 3D reconstruction of them is not only important in theory, but also valuable in practice.This dissertation is focused on camera calibration and 3D reconstruction based on images from structured scenes. The dissertation is organized as follows:Firstly, several methods of straight line detector from single image are discussed and their measuring precisions are compared. A method based on RANSAC algorithm to estimate the vanishing points of three orthogonal images is discussed. Compared with classical least square method the proposed method is better in general or slightly degenerate conditions. Especially in degenerate conditions, its precision is notablely better than traditional method. As a statistic method, its arithmetic speed is limited by sample size and arithmetic complexity which cause low efficiency. But when it's combined with straight line detector, a modification method can be obtained with limited sample size and high arithmetic precision.Secondly, the concept that the three-parameter-camera's projection matrix can be linearly gained from a set of three orthogonal vanishing points is discussed. It is also testified that the projection matrix is uniquely determined under a given world coordinate system. when we have multiple views of the scene, the scalars corresponding to the projection matrices of these views must be consistent so as to obtain a global reconstruction consistently. After retrieving the projection matrices, one may compute the 3D structure of any pair of correspondence that are tracked automatically or given interactively. Most structured object usually composed of many pieces of planar surfaces, thus the structure is easy to obtain with minimal human interactions of the corners of surfaces across two or more views. There is no influence whether vanishing points lie in or out of the image. If vanishing points tend to be in infinite distance, the accuracy of this method will be severely influenced or even ineffective. The method is easy to implement and more accurate and robust results are expected. The result of this method can be used as the initial to do further calibration and high accuracy results will be obtained through further human interaction.a discussion between the use of homography should resume scene specific geometric measurement method. The method of distance of two points in the reference plane ,objects height, the distance of two points in the vertical plane was introduced. These specific geometric measure is useful in the structured scene reconstruction . through the geometric method of the structure scene, some distance of the space points can be calculated. If vanishing points of vertically the reference plane is known, the camera calibration and scene reconstruction through a single image can be completed. Through the restoration of the geometric absolute metric, we can get the relative position and geometric scale between objects in the structured scene. Then the structured scene reconstruction is complete .At last, several sets of real images are tested on the proposed method and the structure design of the 3D reconstruction software is explained. One groups test images are downloaded from the Visual Geometry Group of the University of Oxford, another groups is shooting by me. Canny edge detector and orthogonal regression algorithm are applied to fit the line segments in images, and the vanishing points from the parallel lines are computed using maximum likelihood estimation. The proposed algorithm is used to compute the projection matrices, and reconstruct the object whose corner points of each planar surface are interactively selected. In the end, reconstruction model is obtained through different views. Experiments show that the method is simple and more accurate and realistic models are expected. The method avoids the bottleneck problem of image matching and is easy to implement. It is suitable to reconstruct photo-realistic and accurate model of structured scenes, especially those contain three dominate directions.